融合高频信息增强和注意力机制的视线估计方法  

A Gaze Estimation Method Integrating High⁃Frequency Information Enhancement and Attention Mechanism

作　　者：印洁 沈文忠 邵洁 YIN Jie;SHEN Wenzhong;SHAO Jie(School of Electronics and Information Engineering,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]上海电力大学电子与信息工程学院,上海200090

出　　处：《上海电力大学学报》2024年第3期279-284,共6页Journal of Shanghai University of Electric Power

摘　　要：准确的视线注视方向估计是人机交互和虚拟现实等应用场景中的关键技术。基于外观的视线估计是目前的主流方法,然而,因为眼睛外观、光线条件和头部姿态的多样性,所以无约束环境下的视线估计仍然是一个具有挑战性的任务。提出了一种高频信息视线估计网络(HFA-Net)。首先,在神经网络中加入高频信息提取模块和卷积注意力模块(CBAM),帮助网络减少冗余信息的影响;其次,将视线分为两个角度分别进行回归,并使用独立损失函数进行优化;最后,在公开数据集MPIIGaze上进行训练和测试。实验结果表明,该方法在MPIIGaze上取得了4.17°的最佳角度估计误差,超越目前主流算法。Accurate gaze direction estimation is a crucial technology in various applications such as human-computer interaction and virtual reality.The appearance-based gaze estimation is currently the mainstream method.However,due to the diversity in eye appearance,lighting conditions,and head poses,gaze estimation in unconstrained environments remains a challenging task.This paper proposes a high-frequency amplifier gaze estimation network(HFA-Net).Initially incorporating a high-frequency information extraction module and Convolutional Block Attention Module(CBAM)into the neural network helps reduce the impact of redundant information.Finally,the network divides gaze angles into two separate regression tasks,each optimized with independent loss functions.The proposed method is trained and tested on the publicly available MPIIGaze dataset,and experimental results demonstrate that this approach achieves the best angle estimation error of 4.17°on MPIIGaze,surpassing the current mainstream algorithms.

关 键 词：视线估计 高频信息提取 扩张卷积 卷积注意力模块 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]